Journal of Naval Science and Engineering

2016, Vol. 12, No.1, pp.19-32

Endüstri Mühendisliği/Industrial Engineering

A RICH MIN-MAX VEHICLE ROUTING PROBLEM

Ertan YAKICI
Industrial Engineering Department,
Turkish Naval Academy, Naval Sciences and Engineering Institute, Tuzla, Istanbul
eyakici@dho.edu.tr

Date of Receive: 14.01.2016 Date of Acceptance: 28.03.2016

ABSTRACT
We present a new variant of vehicle routing problem with a min-max objective
function. The problem has different types of service demands satisfied by a heterogeneous
fleet of vehicles. Unlimited service capacitated vehicles serve the demand points with
multiple time windows and requirement of operation synchronization when demand is split
between vehicles. A mixed integer linear programming based heuristic solution approach is
proposed and a numerical study is carried out to assess the performance of the proposed
method.
ÖZ
Bu çalışmada min-max amaç fonksiyonlu yeni bir çeşit araç rotalama problemi
sunulmaktadır. Problemde farklı tipte araçlardan oluşan bir filo ile farklı türde hizmet
talepleri karşılanmaktadır. Hizmet kapasitesi sınırsız olan araçlar, birden fazla zaman
pencereli ve talebin araçlar arasında bölünerek karşılanması halinde operasyon
senkronizasyonu gerektiren talep noktalarına hizmet sağlamaktadır. Problemin çözümü
için tamsayılı doğrusal programlama tabanlı bir sezgisel algoritma yaklaşımı önerilmiş ve
önerilen metodun persformansını değerlendirmek için sayısal bir çalışma yapılmıştır.
Keywords: Vehicle Routing Problem; Heuristic Method; Synchronization.
Anahtar Kelimeler: Araç Rotalama Problemi; Sezgisel Yöntemler; Senkronizasyon.

1. INTRODUCTION

Since the first introduction of Vehicle Routing Problem (VRP) by

Dantzig and Ramser [1], a huge literature is created on VRP and its several
extensions. Like different constraint environments, objective environment
also may have different focuses like minimization of the maximum or
average arrival time. Those objective functions is often seen in the nonprofit
environments like military operations, disaster relief and humanitarian

19
 Ertan YAKICI

logistics (Duran et al. [2], Yakıcı and Karasakal [3], Renkli and Duran [4]).
Naval operations like logistic support of warfare fleets, mine sweeping or
hunting operations is the basic motivation of this research. Type of logistic
support or operation, platform eligibility and other characteristics like
operation speed (or economic speed), desired support or operation periods
comprise several restrictions on the problem.

In the research problem presented in this work, there is a fleet that

consists of vehicles having varying operational capabilities. Some vehicles
are capable of conducting one of the service types, while others can operate
in multiple types of service. Similarly, vehicles have different operation and
transfer speeds. It is assumed that, service demands can be satisfied by one
or more vehicles, demand points have multiple time windows denoting
allowed periods for vehicles to arrive and all of the service demanded by
each demand location should be satisfied by vehicle/vehicles arriving at the
same time window. To the best of our knowledge, although there are very
close problems, our problem is never introduced before in the literature.

A closely related problem is introduced in the work of Yakıcı and

Karasakal [3] where various type of customer demands are satisfied by a
heterogeneous fleet and split delivery is allowed. However, there are
important differences between this problem and our research problem. In the
problem presented in Yakıcı and Karasakal [3], the demand points are
grouped in regions and the vehicles are not allowed to change their regions
once they are assigned, while there is no such restriction in our problem.
The other difference which makes our problem more generalized and
complicated is the existence of realistic operational limitations, namely
multiple time windows and operation synchronization, which are not
handled in Yakıcı and Karasakal [3].

Temporal constraints are encountered in rich VRP problems

frequently. However, so far, the interdependency of vehicles attracted less
attention. The interdependency between the vehicles in operation
synchronization is tedious to handle, especially in developing heuristic
approaches. Several authors try to overcome these issue by allowing

20
 A Rich Min-Max Vehicle Routing Problem

infeasible solutions, evaluating the cost approximately and using indirect

search (Drexl [5]).

The closest solution technique to the one we adopt is presented in the

research article of Bredström and Rönnqvist [6]. They introduce a combined
VRP and scheduling problem with time windows and temporal constraints
and propose an optimization based heuristic to solve real instances. In order
to have a small branch and bound tree they restrict the problem significantly
in number of variables and constraints, then they try to improve the best
known feasible solution.

The heuristic approach we propose is based on the idea of sequential

insertion of vertices and route construction by restricted model solved with a
standard ILP solver.

In the following sections, the problem is introduced with its

mathematical formulation, the proposed heuristic approach is described and
the results of numerical experiments are reported. Finally, conclusion is
given in the last section.

2. PROBLEM DEFINITION

In this section, we present the notation and the definition of our

problem. The total of transfer and service time of a vehicle leaving central
vertex and visiting a given sequence of demand points is referred as “travel
time”. Note that since the operation is considered to finish with the
satisfaction of all service demands, time elapsed in returning to central
vertex is not included in the travel time. Therefore, distance from any vertex
to central vertex is considered to be zero.

The following sets and parameters are used: K = (1,...,|K|): set of

uncapacitated vehicles. G = (V,E): complete directed graph with set of
vertices V = (1,...,n) and set of arcs E consisting of arcs (i,j) where i and j
V. P = (pij): distance matrix between all pairs of vertices in V. S = (sk):
vector of transfer speeds of vehicles in K. Q = (qi): vector of service

21
 Ertan YAKICI

requirements demanded by vertices in V. D = (1,...,|D|): set of service types.

R = (rkd): matrix of service rates (delivered service in unit time) of the
vehicles in K for type of demands in D. Kd: set of uncapacitated vehicles
which can serve d type of service. Vd: set of vertices which demand d type
of service. Wi: set of time windows for vertex i. uiw : Latest time of time
window w for any vehicle to arrive at vertex i. liw : Earliest time of time
window w for any vehicle to arrive at vertex i. M: a big number. : a small
number.

In order to formulate the problem as an Mixed Integer Program

(MIP), let and denote the set of real numbers and {0, 1}, respectively.
The decision variables are described as follows: takes value 1 if
th
edge (i,j) belongs to the k route; 0 otherwise. denotes the
maximum travel time. denotes the service time of vehicle k at vertex
i. denotes the arrival time of vehicle k at vertex i. takes
value 1 if vehicle k arrives at vertex i at time window w. takes
value 1 if vertex i is visited at time window w. Only one time window is
allowed for each vertex in the context of operation synchronization.
is a dummy variable used for subtour elimination.

(1)

(2)

(3)

(4)

(5)

(6)

22
 A Rich Min-Max Vehicle Routing Problem

(7)

(8)

(9)

, (10)

(11)

(12)

(13)

The objective function (1) minimizes the maximum travel time.

Constraint (2) enforces the objective function value to be equal or greater
than the largest travel time. Constraint set (3) ensures that all of the
demands are satisfied. Constraints (4-6) ensure that each vehicle is assigned
to at most only one route and each route starts and ends at central vertex
without multiple visits to vertices. Constraint (7) connects the decision
variables related to routing and servicing by ensuring that the vehicles
should visit the vertices where they serve. Constraints (8) and (9) enforce
the arrivals to occur within time windows. Constraint (10) enforce each
arrival satisfies the elapsed time by previous travel. Constraint (11) connects
the decision variables related to routing and time windows. Constraint (12)
and (13) serve to ensure the operation synchronization by enforcing each
vertex can be visited by the vehicles in only one of its time windows.

23
 Ertan YAKICI

Note that still we can reduce the size of feasible space by demand
type cuts (14) which prevents the vehicles from visiting the demand points
that they cannot serve.

(14)

3. HEURISTIC SOLUTION APPROACH

The heuristic approach is based on the idea of restriction of the MIP model.
Instead of solving the problem as a whole, we solve the problem in steps. At
each step, a group of vertices are added to the model until all of the vertices
are covered. The restricted problem is given a certain period of time for each
step. The traversed arcs of the best solution found is carried to the next step.
Each vehicle can break at most a limited number of arcs from its own route
where some (or all) of the carried arcs are allowed to be broken in order to
visit the vertices added. Another restriction is applied once in a while after a
certain number of steps are taken in between. This restriction fixes all of the
previously traversed arcs and service that are already processed in the
visited vertices. Therefore, the problem restarts again with the remaining
vertices. We call these steps “fixing steps”. After a fixing step, a vehicle
starts from the vertex it visits just before returning to the central vertex and
from the time it finishes its service in that last visited vertex. In order to
apply the restrictions, the following constraints are added to the model.

(15)

(16)

(17)

(18)

The binary parameters fkij, akij and nonnegative parameter Ak level

respectively denote whether the arc is traversed at the previous step (= 1),

24
 A Rich Min-Max Vehicle Routing Problem

whether it is allowed to be broken (= 1) at the current step and the time

when vehicle k completes its job just before returning to central vertex in
the last fixing step. The binary decision variable ekij takes positive value if
the arc (i,j) is broken by vehicle k. The last constraint (18) forces vehicle k
to start at the time given by Ak level after last fixing step. Note that, distance
parameter in the model has to take an additional index indicating the vehicle
it belongs to, and distances from new starting vertices of vehicles to all of
the vertices not covered yet, pk0j, have to be recalculated at fixing steps.

We define two criteria to choose vertices to be added. Given the

worst scenario, the vehicles do not break any arc and continue from the
latest visited vertex just before returning to the central vertex, one of them is
the minimum required time to begin to serve newly added vertex, minRT(i,k),
and the other is the maximum time that can be spent before making an
attempt (leave the current vertex) to visit newly added vertex, maxTS(i,k),
where i V, k Ki and Ki is the set of eligible vehicle set for vertex i. These
values are affected by both of the time windows of the vertex to be added
and the time required by vehicle to reach to that vertex. For a vertex to take
priority in addition to the model, it is better to have less value from both of
these criteria. Note that each of the criteria can be defined for each of the
combinations made by a vertex and a vehicle that are eligible to visit that
vertex. Therefore, while considering all of those combinations, it is also
useful to consider the combinations that give the extreme values: the
minimum value coming from the first criterion and the maximum value
coming from the second one, namely, and
. We define a simple function that determines the
relative priorities of the vertices to be added to the model. There are four
components in the function f(i,Ki): the summation of values for all eligible
vehicles (for first and second criteria) and the extreme values which are
defined in the previous paragraph (for first and second criteria). Those four
values for each vertex are separately scaled between 0 and 1. Therefore,
each vertex has four values between 0 and 1. The summation of those four
values, equally or differently weighted, defines the priority of vertex.

25
 Ertan YAKICI

In the selection of the arcs that are allowed to be broken, the number
of consecutive steps that an arc remain in the solutions is taken as the basic
parameter. If this parameter is set to “0”, all traversed arcs are allowed to be
broken in the next step. As the parameter increases, the number of arcs
allowed to be broken decreases. In all cases, the last arc of all vehicles
which is returning to central vertex is kept breakable. We accept that it is
hard to give a specific comment on the best combination for the remaining
parameters which are the number of added vertices at each step, the number
of steps between fixing steps and the time allowed for solving the restricted
problem. These parameters can be determined with respect to the problem
size and the time given to have a feasible solution.

Since we solve the problem in a stepwise manner and establish a

priority value for the vertices for being added to the model, it is possible to
have infeasibility in subproblems because of time window restrictions of the
vertices. In order to handle this case, we add penalty component to the
objective function. In order the penalty mechanism to work, a new decision
variable is introduced into the model. The objective function (1) and the
constraint (3) is changed as indicated in the expressions (19) and (20).

(19)

(20)

The new binary decision variable penaltyi takes positive value if

adding vertex i is infeasible or it takes much time than allowed to find a
solution including that vertex at the current step. If there is a penalty, the
algorithm turns back to the last fixing step to repair the solution to have it
become feasible. In repairing, fixed solution is released in order to have the
penalized vertices be introduced to the model, where all traversed arcs are
kept fixed but allowing that limited number of arcs from each route can be
broken just as it is applied in constructing steps. The formal algorithm of
construction heuristic is given below. Since repairing procedure is

26
 A Rich Min-Max Vehicle Routing Problem

straightforward, we do not denote repairing steps separately in the

algorithm.

• Step 1. Set heuristic parameters and construction model (objective function

19 subject to constraints 2, 4-13, 15-18, 20), calculate vertex priority values
and create an ordered list of vertices L w.r.t. addition-to-model priority. Go
to Step 2.

• Step 2. If L is not empty, add first nstep vertices (or add all vertices if there
are less than nstep) from L to model and update L, excluding the vertices
added to model, go to Step 3. Otherwise stop.

• Step 3. Solve model in given time period,

– Step 3.1. If this is not a fixing step and all vertices are served
without penalty, set the values to 1 for all traversed arcs (i,j) E
and k K where is equal to 1. Set to 1 w.r.t. given setting
in order to determine the arcs that can be broken, go to Step 2.

– Step 3.2. If this is a fixing step and all vertices are served without
penalty, go to Step 4.

– Step 3.3. If all vertices cannot be served, turn back to the last
fixing step and apply repair procedure. If solution cannot be repaired
go to previous fixing steps to apply repair procedure until the
penalized vertices are added to the model. Update L, go to Step 2.

• Step 4. Exclude all added vertices from model. For all excluded vertices;
keep the traversed arcs, served time windows and serving duration in the
vertices. Update the distances from central vertex to other vertices: For all
vehicles k K, set to where c indicates the last visited vertex of k.
Set the current vertices to “0”, and to for all vehicles
k K. Go to Step 2.

27
 Ertan YAKICI

The solution constructed as it is explained above may have still room

to be improved. One way is improvement within each route and the other is
combining more than one routes to have an opportunity to improve. Before
combining the routes, we choose to check if there is an improvement room
within each route applying a simplified version of the model with
considered vehicle and vertices it has served. Then, we expand this check by
adding two routes, one from the set of longest routes and one from the set of
shortest routes, to the model. In both improvement phases, the traversed arcs
are all allowed to be broken, which means the exclusion of the constraints
(15-17) from the model. Also allowed time window of each vertex that are
served by other vehicles (the vehicles belong to the routes that are not
considered at the current improvement step) are fixed by adding constraint
(21) into the improvement model.

(21)

The binary parameter twniw level is set to positive value, if the vertex i
is served in time window w by any other vehicle that is not considered at the
current improvement step. While choosing the longest and shortest route,
the combination, giving the greatest number of vertices in longest route that
can be served by the vehicle of shortest route, is given priority. While the
set of longest routes consist of only the routes that have the maximum time
length, the set of shortest routes are populated starting from the route(s)
having shortest time length and adding one by one until the set satisfies that
the vehicles included are eligible to serve all vertices served by the set of
longest routes. To complete the improvement phase within reasonable time
period, each improvement iteration is given a certain period of time and
after allowing a certain number of steps (or time), when there is a
deterioration or no-improvement in the objective function, the algorithm is
stopped. The steps of the procedure applied for the improvement phase of
the heuristic approach is given below.

• Step 1. Set improvement model (objective function 1 subject to constraints

2-13, 21). Solve separately in given time periods for each vehicle k K.
Update the solution. Go to Step 2.

28
 A Rich Min-Max Vehicle Routing Problem

• Step 2. Update the set of longest routes Rl and set of shortest routes Rs such
that Rl covers all rl longest routes having equal values, and Rs covers the
shortest rs routes of which vehicles can serve all of the vertices covered in
Rl. Go to Step 3.

• Step 3. If the stopping criterion does not hold, solve improvement model
with the combination of one route each from Rl and Rs, such that the number
of vertices in the longest route that can be served by the vehicle of the short
route be maximized (if there is a tie, choose the combination with shorter
route, if the tie is not broken, break randomly). Update the solution, and go
to Step 2. Otherwise stop.

4. EXPERIMENTS

Since there is no test instances that can serve as benchmark either as

a lower bound or optimal solution to our problem, we choose our test
problems from the test problems given in Yakıcı and Karasakal [3]. We also
applied the same vehicle parameters as they are applied in Yakıcı and
Karasakal [3]. While keeping all of the data of test problems given in Yakıcı
and Karasakal [3] as they are, since it is required in our problem, we add
parameters for two time windows for each vertex. Three different time
window pairs, (0-15, 20-35; 5-20, 25-40; 10-25, 30-45), are determined and
one of these pairs is assigned to each one of vertices except central vertex.
In time window assignment to vertices, we follow the order of given
numbers to vertices and the order of time windows given in the previous
sentence.

Performance of the proposed method is compared to exact solution

method provided by CPLEX 12.6.2.0 ILP solver. In addition, since even the
small instances cannot be solved by exact method, in order to have another
benchmark, we relaxed the time windows such that the first time window of
each node has no upper bound. The lower bounds and best solutions
achieved in twelve hours are reported along with the heuristic methods
performance. In application of heuristic method, at most two minutes (and
10% relative gap criterion, if it is reached earlier) is given for solving the

29
 Ertan YAKICI

restricted MIP at each step and maximum ten minutes is given in

improvement phase. Except 31-node instance, all instances used up ten-
minute improvement period. Improvement phase is only applied after the
completion of construction and stepwise approach is not used in this phase.
In the construction phase, five nodes are added to the model at each step
using equal weight priority function values and at each step only one arc
from each route is allowed to be broken. When not all arcs, but only older
arcs (the arcs which is traversed in the same route consequently at more than
one in the last steps) are defined as breakable, the last arc returning to
central vertex is always allowed to be broken.

Experiments are conducted with a personal computer with 2.6 GHz

CPU and 8 GB RAM. The result of the experiments is reported in Table 1,
where instance identity (ID) is given in the first column. In instance ID, the
first section before the letter “k” denotes the test problem group and number
of vertices as denoted in Yakıcı and Karasakal [3]. Vehicle quantities from
each type are indicated after the letter “k”. Next four columns are related to
the solutions of MIP and relaxed MIP. MIP refers to the model defined by
objective and constraints (1-13), while Relaxed MIP refers to the same
model where only first time windows are applied without upper bound. For
MIP and Relaxed MIP, the lower bound and the incumbent solution from
exact solution method is reported after twelve-hour period is elapsed. In the
columns under the “Heuristic” title, solution for the problem with original
time windows (not relaxed), time elapsed in the construction phase of
solution process (in minutes and seconds) and parameter settings are given.
In the “Settings” column, first field (before slash) denotes the number of
steps that an arc should remain in the solution to be breakable and second
field denotes the required number of vertices to be covered to apply fixing
steps. As presented in this column, we start fixing when the step with 15
vertices is solved. Then, at every increment of 10 more vertices in the
problem, fixing step is applied.

30
 A Rich Min-Max Vehicle Routing Problem

Table 1. Experiment Results

MIP Relaxed MIP Heuristic
Instance ID LB Sol'n LB Sol'n Sol'n Time Setting
E031-k1/1/2/4 2.33 - 6.49 31.53 35.28 7'23" 0/15-25
E031-k1/1/2/4 2.33 - 6.49 31.53 36.53 6'57" 1/15-25
P-n51-k1/1/2/4 2.29 - 16.99 - 54.28 10'30" 0/15-25-…-45
P-n51-k1/1/2/4 2.29 - 16.99 - 58.10 10'22" 1/15-25-…-45
E-n76-k1/1/2/4 2.18 - 13.74 - 89.05 15'04" 0/15-25-…-65
E-n76-k1/1/2/4 2.18 - 13.74 - 91.04 14'15" 1/15-25-…-65

Observing Table 1 yields that we have rather poor lower bounds and
no solutions at all from commercial solver. When the problem is relaxed and
solved by the same solver, we can find better lower bounds which serves as
lower bounds to the original problem (MIP). Since the solver does not
perform well in finding solutions, we believe these lower bounds are still far
away from optimal solutions. Therefore, while we cannot determine the
quality of heuristic solution, we can claim that the heuristic is useful in this
case when there is no other algorithm that can find better solutions.

5. CONCLUSION

We introduce a rich VRP in which there is a heterogeneous fleet

having varying operational capabilities and it is assumed that, service
demand of a demand point can be satisfied by one or more vehicles in one
of assigned multiple time windows.

An ILP formulation of the problem is given and an ILP based

heuristic is developed for solving the problem. We have employed and
tailored three small test problems from an earlier study [3] about a similar
problem having no time window and synchronization restrictions. Even with
these small instances, the commercial solver cannot find any feasible
solution in twelve hours while heuristic solution is able to find solutions in
less than half of an hour for the largest instance. Also lower bounds found
by the exact method solver in twelve hours are very poor. Therefore, we do

31
 Ertan YAKICI

not have an idea about the solution quality, however we consider that the
proposed method is useful in the absence of a better solution.

REFERENCES

[1] Dantzig G.B., and Ramser J,H., (1959). The truck dispatching
problem, Management Science, vol. 6(1), pp. 80–91.
[2] Duran S., Gutierrez M.A., and Keskinocak P., (2011). Pre-
positioning of emergency items for care international, Interfaces, vol. 41(3),
pp. 223–237.
[3] Yakıcı E., and Karasakal, O., (2013). A minmax vehicle routing
problem with split delivery and heterogeneous demand, Optimization
Letters, vol. 7(7), pp. 1611–1625.
[4] Renkli Ç., and Duran S., (2015). Pre-positioning disaster response
facilities and relief items, Human and Ecological Risk Assessment: An
International Journal, vol. 21(5), pp. 1169–1185,.
[5] Drexl M., (2012). Synchronization in vehicle routing - a survey of
VRPs with multiple synchronization constraints, Transportation Science,
vol. 46(3), pp. 297–316.
[6] Bredström D., and Rönnqvist M., (2008). Combined vehicle routing
and scheduling with temporal precedence and synchronization constraints,
European Journal of Operational Research, vol. 191(1), pp. 19–31.

32